What does fanning out 100 AI subagents cost? A calculation from measured runs

100 small Sonnet calls cost $0.36 at the observed median. 100 full Crewshal attempts cost $280.72 at the observed mean. Both are calculations at list price. We did not start parallel subagents.

The small-call estimate uses 15 calls across five tasks. Sonnet passed 12/15 strictly (Wilson 95% interval 54.8% to 93.0%). Per-call cost ranged from $0.0034 to $0.0102. This is cost per call, not cost per correct answer.

The worker estimate uses all 33 attempts, including failures. Per-attempt cost ranged from $0.89 to $4.83. Neither calculation predicts the bill for a new task set.

What we priced

A unit here is a separate CLI process or a full Crewshal worker attempt. We did not measure Claude Code's built-in subagents within one session.

We used three workloads:

1. Minimal replies through each CLI, one at a time.
2. Five short tasks with strict validators, through selected CLI configurations.
3. Full Crewshal attempts on SWE-bench Verified, with onboarding, planning, editing, verification and review.

The calls used subscriptions. Every dollar figure is a list-price calculation, not an invoice. Counts of 10, 100 and 500 are assumed scenario sizes, not measured fan-outs.

Minimal replies: input and cache scenarios

Claude Code used Haiku 4.5 for five minimal replies. Its median input count was 6,761 tokens, including cache reads and writes. The public aggregate has no token range. No empty-prompt control isolates the CLI prefix or establishes tokens sent before work.

Codex also made five minimal replies, using its default model. The corrected public aggregate reports 17,051 inclusive input tokens, including 13,184 cache-read tokens. Cached tokens are a subset and must not be added again. We retain no Codex input or cost projection: different models and a minimal reply do not isolate a common startup prefix.

We repriced the Claude input at Sonnet 5.5 rates per million tokens: $2 input, $0.20 cache read and $4 one-hour cache write. This assumes the same input size. It excludes output and orchestration costs.

Each cell is a calculation. These are cache assumptions, not measured bounds. The first scenario assumes a write finishes before later calls start and every prefix matches.

As a calculation from reported first-turn counters, Claude first-turn read shares averaged 18.7% (n = 6; range 18.680% to 18.689%). Codex averaged 55.5% (n = 3; range 47.7% to 59.3%). These are means of process shares, not pooled token shares. Claude used a shortened ledger; Codex used the earlier longer ledger. The routes are not matched.

In later Claude processes, first-turn cache reads exceeded writes in 0/4 cases (Wilson 95% rate interval 0% to 49.0%). This is a counter proxy. It cannot show which ledger tokens a process read.

Small calls: cost and strict passes

Each selected configuration made 15 calls across five tasks. Costs include all calls, including strict validator misses. Every dollar value below is a calculation from reported tokens at list price.

The cost ranges and pass intervals overlap. This sample does not rank these configurations. The task set is easy and selected. Configurations with no misses hit a ceiling. Sonnet's three misses gave the correct final number but added working lines that the exact-text validator rejects.

Each small call includes its own CLI prompt at that run's size. Do not add the minimal-reply scenario to it. Models, flags and prompts differ across runs. We did not test the cause of cost differences.

Full worker attempts

Crewshal made 33 attempts across two campaigns and two platform builds. It resolved 25; failures stay in cost totals. Mean notional cost was $2.81 per attempt, with an observed range of $0.89 to $4.83. These totals include compaction calls.

Scaled ranges are scenario spreads, not confidence intervals. The campaigns include replacement instances. New tasks can require different tokens and have different outcomes.

Duration: serial scenarios, no parallel bound

N × median duration is a serial scenario. It is not the measured sum or expected total. A one-unit median is a reference time, not a parallel lower bound.

Ranges are not confidence intervals. The minimal-reply routes used different models. All timings include the answer and process exit or worker completion.

Simultaneous completion depends on the maximum duration, not the median. We did not measure parallel completion, rate limits, machine limits or contention. Other work shared the Mac and subscription accounts.

Evidence limits

Pre-call protocol declaration is unverified from the selected public extracts. File times and final protocol text do not establish what was declared before the calls. This calculation adds no new experiment or advance-declaration proof.

The linked minimal-reply, head-to-head and cache extracts retain final protocol text with advance-declaration headings. No preserved pre-call protocol versions were supplied for these extracts. Those headings do not prove what existed before the calls. Historical raw records and review verdicts stay unchanged.

The public-copy note describes a retained final summary and states that pre-call versions are not preserved in that summary. This corrects the label; it does not verify advance declaration. The corrected labels need a fresh source and text recheck before publication.

Use these calculations as scenarios. Measure your own tasks before using them as a budget.

Disclosure: I build Crewshal. The full-attempt numbers come from Crewshal's own runs.